CN114556235A

CN114556235A - Diagnostic device

Info

Publication number: CN114556235A
Application number: CN202080072563.5A
Authority: CN
Inventors: 托尔斯腾·哈弗李宁-尼尔森; 玛丽亚塔·皮罗宁; 利兹·约恩苏; 维萨-马蒂·蒂卡拉
Original assignee: Kemira Oyj
Current assignee: Kemira Oyj
Priority date: 2019-10-16
Filing date: 2020-10-13
Publication date: 2022-05-27
Also published as: AU2020367256A1; US20240103507A1; KR20220084069A; FI20195892A1; EP4046079A1; BR112022006106A2; CA3152883A1; WO2021074490A1

Abstract

The present invention provides a diagnostic device that utilizes pre-processed measurement data, ML values, and interpreted values. By using all these values/data, phenomena, events and behaviour of the process can be analyzed, whereby many aspects can be taken into account.

Description

Diagnostic device

Technical Field

The invention relates to diagnosis of a processor and control thereof. Such as water treatment equipment, paper machines, etc.

Background

Today, machine learning algorithms are used with systems that analyze and estimate the behavior of processes such as paper machines or water treatment. The processes are typically multivariate processes, so they are difficult to track or understand. Machine learning provides the system with the ability to learn automatically and to improve from experience without explicit programming. Thus, the computer system uses Machine Learning (ML) utility algorithms and statistical models to perform a particular task or tasks without the use of explicit instructions. There are several ML algorithms. Only some of which are mentioned here: linear regression, logistic regression, K-means, feed forward neural networks, and the like.

Reasoning about the results of the ML algorithm is often difficult to interpret, especially from complex processing. Thus, the interpretation value is used to help the user interpret the result. Therefore, the interpretation values are used to interpret how the ML algorithm yields a particular result, and also to classify the workings of the process. The explained value is obtained by using, for example, a SHAP (xiapril addition explanation) value, a LIME method, or a deep lift method.

Fig. 1 shows an example of a known control device, wherein a process 1 is driven by an actuator 2, the actuator 2 being controlled by a controller 3. The measured values 4 are obtained from this process and they are used as feedback data for the controller. The controller compares the measured values with one or more set point values 5 and forms control commands for the actuators.

The measurement 4 can also be used for other purposes, in which case it is convenient to pre-process 6 the measurement data before actual use. For example, preprocessing may include data merging, aligning time formats, modifying metadata, data validation, and the like. In the example of fig. 1, Machine Learning (ML)7 is used to extract information and patterns in large datasets. Matching learning algorithms are typically based on statistical models, and computers can perform specific tasks without exact instructions, but rely on recognizing patterns. The identified pattern may be obtained by building a mathematical model based on a training data set. Prediction (simulation) and pattern recognition can be performed by inputting new data into the mathematical model.

Because it is difficult to see what happens in the process from the output of the ML (prediction/simulation), the interpretation value of 8, SHAP value in the embodiment of FIG. 1, is used to track how the ML predicts links back to the input variables. For each prediction, a number of scoring stages is calculated for each input variable to indicate the contribution of that variable to the final prediction. These rating numbers may be considered as interpretation values indicating the importance of the input value at a given point in time.

The interpretation values are used to verify how the ML algorithm and ML model work 9. This can be done more easily from interpreted values than from ML predictions. Therefore, if the ML model does not work properly, it can be modified.

The ML value is used in the prediction unit 10 to predict the behavior of the process. The prediction may be used to provide a recommendation 12 to process 1. The prediction can also be used to suggest a correction 11 to change the set point 5 of the controller 3.

Although ML values are used, without means it can also utilize other data in an automated way.

Disclosure of Invention

It is an object of the invention to provide a diagnostic device which makes use of preprocessed measurement data, ML values and interpretation values. By using all these values/data, phenomena, events and behaviour of the process can be analysed, so that many aspects can be taken into account. This can be done automatically. This object is achieved in the manner described in the independent claims. The dependent claims describe different embodiments of the invention.

The inventive diagnostic device for multivariate processing comprises a data processing module 6 for processing measurement data of the multivariate process and performing a pre-processing of the measurement data 6A. The apparatus further comprises a machine learning module 7 to perform machine learning of the values 7A from the pre-processed measurement data 6A. The diagnostic device further comprises an interpretation value module 8 for forming an interpretation value 8A from the machine-learned value 7A and a deviation calculation module 14. The deviation calculation module is arranged to calculate a deviation 8D between the interpreted value 8A and the normal interpreted value 8N, a deviation 7D between the machine learned value 7A and the normal machine learned value 7N, and a deviation 6D between the pre-processed measurement data 6A and the normal pre-processed measurement data 6N. The diagnostic apparatus further comprises at least one estimator 15, each estimator being arranged to track a specific interference condition of the multivariate process using said

deviations

6D, 7D, 8D and to form an estimate 33 of the severity of the interference condition.

Drawings

The invention is described in more detail hereinafter with reference to the accompanying drawings, in which

Figure 1 illustrates an example of a prior art device,

figure 2 illustrates an example of a diagnostic device according to the invention,

figure 3 illustrates an example of an estimator according to the invention,

figure 4 illustrates another example of an estimator according to the present invention,

FIG. 5 illustrates an example of LE or fuzzy mapping, an

FIG. 6 illustrates other examples of LE or fuzzy mappings.

Detailed Description

Fig. 2 illustrates an example of the inventive diagnostic apparatus for multivariate processing 1. The process may comprise a plurality of processes and thus as a whole it may be a combination of processes running together. It comprises a data processing module 6 for processing the measurement data of the multivariate process and for performing a pre-processing of the measurement data 6A. The apparatus further comprises a machine learning module 7 to perform machine learning of the values 7A from the pre-processed measurement data 6A. The diagnostic device further comprises an interpretation value module 8 for forming an interpretation value 8A from the machine-learned value 7A, and a deviation calculation module 14. The deviation calculation module is arranged for calculating a deviation 8D between the interpreted value 8A and the normal interpreted value 8N, a deviation 7D between the machine learned value 7A and the normal machine learned value 7N, and a deviation 6D between the pre-processed measurement data 6A and the normal pre-processed measurement data 6N. The deviation calculation module may have several modules to perform the calculation, such as a module for calculating a deviation 8D between the interpreted value 8A and the normal interpreted value 8N. The deviation calculation module 14 may also be a distributed module with independent modules for performing the calculations.

The diagnostic apparatus further comprises at least one estimator 15, each estimator being arranged to track a specific disturbance or a specific quality condition of the multivariate process using said

deviations

6D, 7D, 8D and to form an estimate 33 of the severity of the disturbance condition. For example, in papermaking, one estimator may be arranged to track the retention of fine particles, while another estimator may track sizing performance. The output 15A of each estimator 15 may be used as such or together with the outputs of the other estimators to provide recommendation and/or guidance commands 16, such as commands to change the set point of the controller 3, change the recommended processing 1 of raw materials, recommendations to improve water washing, recommendations to optimize retention, quality indicators indicating the health of the process or sub-process, etc. These recommendations can vary depending on the process in question. The output 15A of each estimator can be used alone or in combination with the outputs of the other estimators to control, optimize, or troubleshoot the multi-variable process. Controlling and/or optimizing may include controlling one or more of the amount of chemical dosed, the point of dosing of the chemical, the dosing interval of the chemical, the selection of the type of chemical used in the treatment, and the treatment conditions, such as pH, temperature, flow rate of the treatment stream.

The machine-learned interpretation value and the machine-learned normal interpretation value are, for example, a SHAP value, a value from the LIME method, a value from the DeepLIFT method, or any other possible interpretation value.

The LIME method interprets individual model predictions, which are based on local approximations of the model around a given prediction. The LIME refers to the simplified input x as an interpretable input. The mapping x ═ hx (x) converts the binary vector of the interpretable input into the original input space. Different types of hx-maps are used for different input spaces.

DeepLIFT is a recursive predictive interpretation method. It assigns to each input xi a value C Δ xi Δ y that represents the effect of setting the input to a reference value rather than its original value. This means that the deep lift mapping x ═ hx (x) converts the binary value into the original input, where 1 denotes that the input takes its original value and 0 denotes that the reference value is taken. The reference value represents a typical background value without information of the feature.

The SHAP (SHAPLYAdtive explicit edition) interpretation value attributes each feature to the change predicted by the model when the feature is adjusted. These values explain how the expected value E f (z) to be predicted is obtained from the base value if we do not know any characteristics of the current output f (x). The order in which features are added to the desired values is important. However, this has been taken into account in the SHAP value.

Fig. 2 also shows (also like fig. 1) a process 1 driven by an actuator 2, which is controlled by a controller 3. The measured values 4 are taken from the process and the controller compares the measured values with one or more set point values and forms control commands 3A for the actuator 2.

As already described, the measured values 4 can also be used for other purposes and can be preprocessed 6. For example, preprocessing may include data merging, aligning time formats, modifying metadata, data validation, and the like. In the example of fig. 2, machine learning 7 is used to extract information and patterns in large datasets. The identified pattern may be obtained by building a mathematical model based on a training data set. Prediction (simulation) and pattern recognition can be performed by inputting new data into the mathematical model.

Interpret value 8, like the SHAP value, is typically used to track 9 how the ML value links back to the input variables. For each prediction, a number of scoring stages is calculated for each input variable to indicate the contribution of that variable to the final prediction. These rating numbers are explanatory values representing the degree of importance of the input value at a given time point.

It may be noted that the deviation/error between the normal interpretation value and the current ML prediction/estimation interpretation value, and the deviation between the normal ML value and the ML value, as well as the normal (preprocessed) measurement data and the preprocessed measurement data are calculated. The normal interpretation value may be a repository value found from a good run cycle of the process. Thus, the machine-learned normal interpretation value 8N, the normal machine-learned value 7N, and the normal pre-processing measurement data 6N are values/data 13A that have been derived from the good run cycle of the process. For example, normal values may be derived as simple or intermediate values for these good periods. Normal operation of the process occurs during a time period when the process or combined process is operating well. Thus, for all data (pre-processing, ML prediction and ML interpretation values), a normal (optimal) value can be given or estimated (from the stored values). Thus, there may be a library of normal historical values where it has been determined that the process is operating in an optimal manner.

Thus, during operation where the individual or combined processes do not function optimally, a discrepancy, deviation or error is detected from the measured value, the ML value and the interpreted value. This is detected as a deviation from the normal value. The

differences

6D, 7D, 8D (see fig. 3) from the

normal values

6N, 7N, 8N are used as input to the estimator 15. Although the deviation calculation module 14 is shown as a separate module, it may also be part of the estimator 15. Typically, the deviation is related to the error. The magnitude of the error indicates the need to change the set point value or how much the set point value should be changed.

Fig. 3 shows an example of an estimator 15 which uses the deviations/

errors

6D, 7D, 8D. The example of FIG. 3 shows three error values for three variables, but as shown, more variables and offset values may be used if desired. Thus, at least one error/deviation value 6D, ML of the measurement data 6A at least one error/deviation value 7D of the value 7A and at least one error/deviation value 8D of the interpretation value 8A may be used in the estimator of the invention.

The estimator 15 comprises at least one P-

module

17, 17A, 17C and an I-

module

18, 18A, 18C or a D-

module

19, 19A, or any combination of these modules. As previously mentioned, the offset is the data input into the module. The estimator further comprises an

input mapping module

20, 21, 22, 20A, 21A, 22A, 20C, 21C for each

output

23, 24, 25, 23A, 24A, 25A, 23C, 24C of the module. Further, the estimator comprises a summing module 26 for summing the

outputs

27, 28, 29, 27A, 28A, 29A, 27C, 28C of the

input mapping modules

20, 21, 22, 20A, 21A, 22A, 20C, 21C and an output scaling module 30 for scaling the output 31 of the summing module. Further, the estimator comprises an output mapping module 32 to provide a normalized output 33. The normalized output is an estimate, which is used for recommendations, etc., as described above.

P, I and D-

modules

17, 17A, 17C, 18A, 18C, 19A and combinations thereof PI, PD, ID and PID are known, but no explained value or deviation/error of the ML value has been used as input before. The P-

block

17, 17A, 17C has a weighting coefficient which is multiplied with the input error value. The I-module includes an

integrator unit

118, 118A, 118C that integrates the input error value for a particular period. The integrated input error value is multiplied by a

second weighting factor

180, 180A, 180C. The D-module comprises a

differentiator unit

119, 119A forming a derivative of the error value during a certain period. The derivative is multiplied by a

third weighting factor

190, 190A. It can be seen that all of the P, I and D blocks and their combinations have a weighting factor unit. The units may have the same weighting coefficients or different weighting coefficients. The weighting coefficients may weight the importance of the proportional (P), integral (I) and derivative (D) portions of the error value, and may also adjust or fine tune the estimate by increasing or decreasing the calculated contribution of each single input.

It is not always necessary to have all of the P, I and D modules, but as previously mentioned, they may be in the estimator if they are actually used. In the embodiment of fig. 3, P, I and the D module together provide PID calculations for interpreting error value 8D and ML error value 7D and PI calculations for measuring error 6D of the data.

Thus, the estimator according to the invention comprises at least one module arranged to process the deviation 8D between the interpreted value 8A and the normal interpreted value 8N, at least one module arranged to process the deviation 7D between the machine-learned value 7A and the normal interpreted value 7N, and at least one module arranged to process the deviation 6D between the pre-processed measurement data 6A and the normal pre-processed measurement data 6N. The estimator may use different numbers of inputs (offsets). For example, the estimator may use only one deviation of the measurement data, four deviations of four different ML values, and two deviations of two different interpretation values.

Fig. 4 shows another possible example where the D-block is not needed and therefore the estimator of this example has PI calculations. As described above, the estimator may have only those modules required for P, I, D, PI, PD, ID or PID calculations of an embodiment of the set point controller. It is also worth mentioning that the estimator may perform different calculations for different error values. For example, the embodiment of FIG. 3 may be modified to another scheme, with PID calculation for error value 8D and P calculation for another error value 7D (i.e., I and

D modules

18A and 19A have been deleted).

As described above, the set point estimator further includes an

input mapping module

20, 21, 22, 20A, 21A, 22A, 20C, 21C for each

output

23, 24, 25, 23A, 24A, 25A, 23C, 24C of the P, I and D modules. See fig. 3. The input map converts the result of each output of P, I or the D-module to a value between-2 and 2. This can be seen as a normalization of the values. The input map is formed by Language Equations (LE) or fuzzy logic. Non-linearities can be conveniently taken into account by using input mappings. The adjustment of the estimator is also relatively smooth because the properties of the processing are taken into account in the input mapping. The mapping module of the estimator may use any mapping curve alone. For example, in fig. 3,

modules

20 and 20A may be formed by LEs, or one module 20 may have been formed by LEs while the other module 20A is formed by fuzzy logic.

Fig. 5 shows an example of a mapping curve 50, which is formed by a linguistic equation or fuzzy logic. X is an input variable that is converted to an output variable Y. The maximum and minimum values of X and Y are determined. A linear formula (e.g., Y ═ ax + b) determines that the Y value between X occurs between the maximum and minimum values. If X is greater than the maximum X value, then Y is the maximum Y. If X is less than the minimum X value, then Y is the minimum Y.

The mapping curve may also be another curve than the linear curve. It may be another curve that matches better with the processed features. Fig. 6 shows two other possible examples of mapping curves. The solid line depicts the piecewise-linear mapping curve 60 and the dashed line depicts the S-curve mapping 61. Other curves are possible. Thus, referring to FIG. 3, the mapping module may use any mapping curve alone. For example,

modules

20 and 20A may have the same mapping curve, such as a linear curve, or different curves, such as different linear curves, or a piecewise linear curve and an S-curve.

The

outputs

27, 28, 29, 27A, 28A, 29A, 27C, 28C of the

input mapping modules

20, 21, 22, 20A, 21A, 22A, 20C, 21C are summed in a summing module 26. Thus, all deviation/error values are taken into account. The sum output 31 is then scaled by an output scaling module 30 and the scaled sum is normalized by an output mapping module 32 to provide a normalized output 33, i.e. the estimator output.

Furthermore, the output of one estimator may be used as input to another estimator along with any combination of measurements, ML predictions, and performance values (e.g., SHAP), which provides a cascading connection between the estimators.

The method of the present invention for forming an estimate of the severity of an interference condition or a quality condition in a multivariate process utilizes the diagnostic device described herein for forming an estimate of the severity of an interference condition or a quality condition. The method uses an estimate of the severity of the disturbance condition or quality condition to provide recommendations and/or guidance commands in the multivariate process to control and/or optimize the multivariate process. The controlling and/or optimizing may include controlling one or more of: the amount of chemical added, the point of addition of the chemical, the interval between addition of the chemical, the choice of the type of chemical used in the process, the process conditions, such as pH, temperature, flow rate of the process stream, and process stream delay, such as pulp, or water flow delay in a process equipment, such as a tower, tank, pulper, tank, or other process equipment.

The method of the invention may control an industrial process, such as a multivariable process, such as a pulp process, a paper, cardboard or tissue making process, an industrial water or wastewater treatment process, a raw water treatment process, a water reuse process, a municipal water or wastewater treatment process, a sludge treatment process, a mining process, a petroleum recovery process, or any other industrial process.

As described above, the present invention provides an automated way to provide an estimator for analysis process 1. For example, the treatment may be a water treatment process or a paper making process. The process may be an industrial process, for example: pulp treatment, paper making, cardboard making or tissue making treatment, industrial water or wastewater treatment process, raw water treatment process, water reuse treatment, municipal water or wastewater treatment process, sludge treatment process, mining treatment, petroleum recovery treatment or any other industrial treatment. The process is typically a multivariate process and therefore requires a large number of measurements to be taken. To understand how the ML algorithm reaches the predicted value, an interpretation value is formed to evaluate the input parameters. There are also normal measurement data, ML values and interpretation values that indicate that the process is working well, and the deviation/error values of the values/data can be formed and used for analysis purposes.

The apparatus of the invention may be located at the same location as the tracked process. However, it may also be located in another place, which makes it possible to track the process remotely. For example, the measurement data 4 is sent via a communication network to the inventive diagnostic device, which processes the measurement data and sends an estimator output, which can be used to adjust the processing. The output of the estimator may be sent to the owner of the process, a maintenance center for the process, or any desired destination.

From the above it is obvious that the invention is not limited to the embodiments described herein, but can be implemented with many other different embodiments within the scope of the independent claims.

Claims

1. A diagnostic apparatus for multivariate processing, the apparatus having: a data processing module (6) for processing the measurement data of the multivariate process and performing a pre-processing of the measurement data (6A); and a machine learning module (7) for executing a machine learned value (7A) from the pre-processed measurement data (6A), characterized in that the diagnostic device comprises: an interpretation value module (8) for forming an interpretation value (8A) from the machine-learned value (7A); and a deviation calculation module (14) for calculating a deviation (8D) between the interpreted value (8A) and a normal interpreted value (8N), a deviation (7D) between the machine-learned value (7A) and a normal machine-learned value (7N), and a deviation (6D) between the pre-processed measurement data (6A) and normal pre-processed measurement data (6N),

the diagnostic apparatus further comprises at least one estimator (15), each estimator being arranged to track a particular disturbance condition or quality condition of the multivariate process using the deviation (6D, 7D, 8D) and to form an estimate (32) of the severity of the disturbance condition or the quality condition.

2. The diagnostic device of claim 1, wherein the machine-learned interpretation value and machine-learned normal interpretation value are a SHAP value, a value from a LIME method, a value from a DeepLIFT method, or any other possible interpretation value.

3. The diagnostic device of claim 2, wherein the machine learned normal interpretation values (8N), normal machine learned values (7N) and normal pre-processing measurement data (6N) are values/data (13A) that have been derived from a good run of processing.

4. Diagnostic device according to claim 3, characterized in that the estimator comprises at least one P-module (17, 17A, 17C), I-module (18, 18A, 18C) or D-module (19, 19A, 19C), or any combination of these modules,

at least one module arranged to process a deviation (8D) between said interpreted value (8A) and a normal interpreted value (8N),

at least one module arranged to process a deviation (7D) between the machine-learned value (7A) and a normal machine-learned value (7N),

at least one module arranged to process a deviation (6D) between the pre-processed measurement data (6A) and normal pre-processed measurement data (6N).

5. The diagnostic apparatus of claim 4 wherein the estimator further comprises an input mapping module (20, 21, 22, 20A, 21A, 22A, 20C, 21C) for each output (23, 24, 25, 23A, 24A, 25A, 23C, 24C) of the module, a summing module (26) for summing the outputs (27, 28, 29, 27A, 28A, 29A, 27C, 28C) of the input mapping module, an output scaling module (30) for scaling the output (31) of the summing module, an output mapping module (32) for providing a normalized output (33) as the estimator output.

6. The diagnostic device of claim 5, wherein the mapping module (20, 21, 22, 20A, 21A, 22A, 20C, 21C, 32) is formed by a linguistic equation or fuzzy logic.

7. The diagnostic device according to claim 6, characterized in that the mapping curve of the mapping module (20, 21, 22, 20A, 21A, 22A, 20C, 21C, 32) provides a linear curve, a piecewise linear curve, an S-curve and/or other curve forms.

8. The diagnostic device of any one of claims 1 to 7, characterized in that it comprises at least one deviation calculation module (14) to provide a deviation (6D, 7D, C) between a machine-learned interpretation value (8A) and a machine-learned normal interpretation value (8N),

8D) -a deviation (7D) between the machine-learned value (7A) and the normal machine-learned value (7N), and/or-a deviation (6D) between the pre-processed measurement data (6A) and the normal pre-processed measurement data (6N).

9. The diagnostic device according to claim 8, characterized in that the deviation calculation module (14) is part of the estimator (15).

10. The diagnostic device according to claim 8, characterized in that the deviation calculation module (14) is a module separate from the estimator (15).

11. A diagnostic apparatus as claimed in any one of claims 1 to 10, wherein said estimate (32) of one estimator is an input to another estimator for use by said another estimator.

12. A method for forming an estimate of the severity of an interference condition or a quality condition in a multivariate process, characterized in that a diagnostic apparatus according to any one of claims 1-11 is used to form the estimate of the severity of the interference condition or the quality condition.

13. The method of claim 12, wherein the estimation of the severity of the disturbance or quality condition is used to provide recommendations and/or guidance commands in the multivariate process to control and/or optimize the multivariate process.

14. The method according to claim 12 or 13, wherein the multivariate process is an industrial process, such as a pulp process, a paper making, a cardboard making or tissue making process, an industrial water or wastewater treatment process, a raw water treatment process, a water reuse process, a municipal water or wastewater treatment process, a sludge treatment process, a mining process, a petroleum recovery process or any other industrial process.

15. The method of claim 13, wherein the controlling and/or optimizing comprises controlling one or more of: the amount of chemical added, the point of addition of the chemical, the interval of addition of the chemical, the selection of the type of chemical used in the treatment, the treatment conditions, such as pH, temperature, flow rate of the treatment stream, and treatment stream delay, such as in a treatment apparatus, such as a tower, tank, pulper, tank, or other treatment apparatus, such as pulp, or water flow delay.